brintos

brintos / llvm-project-archived public Read only

0
0
Text · 23.6 KiB · 51bf4a2 Raw
492 lines · plain
1// RUN: mlir-opt --transform-interpreter -split-input-file -verify-diagnostics %s | FileCheck %s2 3#map = affine_map<()[s0] -> (-s0 + 12, 7)>4 5// CHECK-LABEL: @static_sizes_output_divisible6func.func @static_sizes_output_divisible(%arg0: tensor<24x12xf32>,7                                         %arg1: tensor<12x25xf32>,8                                         %arg2: tensor<24x25xf32>,9                                         %iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {10  %0 = affine.min #map()[%iv2]11 12  //      CHECK: %[[T0:.*]] = tensor.extract_slice %13  //      CHECK: %[[T1:.*]] = tensor.extract_slice %14  //      CHECK: %[[T2:.*]] = tensor.extract_slice %15  %1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>16  %2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>17  %3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>18 19  //  CHECK-DAG: %[[CST:.*]] = arith.constant 0.20 21  //      CHECK: %[[T3:.*]] = tensor.pad %[[T0]] nofold22  //      CHECK: tensor.yield %[[CST]]23  //      CHECK: %[[T4:.*]] = tensor.pad %[[T1]] nofold24 25  //      CHECK: %[[T5:.*]] = linalg.matmul26  // CHECK-SAME:              ins(%[[T3]], %[[T4]] : tensor<4x7xf32>, tensor<7x5xf32>)27  // CHECK-SAME:              outs(%[[T2]] : tensor<4x5xf32>)28 29  //      CHECK: %[[T6:.*]] = tensor.extract_slice %[[T5]]30  //      CHECK: %[[T7:.*]] = bufferization.materialize_in_destination %[[T6]] in %[[T2]]31  %4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>32  %5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>33  func.return %5 : tensor<24x25xf32>34}35 36module attributes {transform.with_named_sequence} {37  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {38    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op39    %padded, %pad, %copy_back = transform.structured.pad %0 {40      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],41      padding_dimensions=[0, 1, 2],42      nofold_flags=[1, 1, 0]43    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.op<"bufferization.materialize_in_destination">)44    %p = transform.num_associations %copy_back : (!transform.op<"bufferization.materialize_in_destination">) -> !transform.param<i64>45    // expected-remark @below {{1}}46    transform.debug.emit_param_as_remark %p : !transform.param<i64>47    transform.yield48  }49}50 51// -----52 53#map = affine_map<()[s0] -> (-s0 + 12, 7)>54 55// CHECK-LABEL: @pad_to_multiple56func.func @pad_to_multiple(%arg0: tensor<24x12xf32>,57                           %arg1: tensor<12x25xf32>,58                           %arg2: tensor<24x25xf32>,59                           %iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {60  %0 = affine.min #map()[%iv2]61  %1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>62  %2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>63  %3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>64 65  //      CHECK: linalg.matmul66  // CHECK-SAME:     ins(%{{.*}}, %{{.*}} : tensor<4x7xf32>, tensor<7x6xf32>)67  // CHECK-SAME:     outs(%{{.*}} : tensor<4x6xf32>)68  %4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>69  %5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>70  func.return %5 : tensor<24x25xf32>71}72 73module attributes {transform.with_named_sequence} {74  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {75    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op76    %padded, %pad, %copy_back = transform.structured.pad %0 pad_to_multiple_of [2, 2, 1] {77      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],78      padding_dimensions=[0, 1, 2],79      nofold_flags=[1, 1, 0]80    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)81    transform.yield82  }83}84 85// -----86 87#map = affine_map<()[s0] -> (-s0 + 12, 7)>88 89// CHECK-LABEL: @parametrized_pad_to_multiple90func.func @parametrized_pad_to_multiple(%arg0: tensor<24x12xf32>,91                                        %arg1: tensor<12x25xf32>,92                                        %arg2: tensor<24x25xf32>,93                                        %iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {94  %0 = affine.min #map()[%iv2]95  %1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>96  %2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>97  %3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>98 99  //      CHECK: linalg.matmul100  // CHECK-SAME:     ins(%{{.*}}, %{{.*}} : tensor<4x7xf32>, tensor<7x6xf32>)101  // CHECK-SAME:     outs(%{{.*}} : tensor<4x6xf32>)102  %4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>103  %5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>104  func.return %5 : tensor<24x25xf32>105}106 107 108module attributes {transform.with_named_sequence} {109  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {110    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op111    %c2 = transform.param.constant 2 : i64 -> !transform.param<i64>112    %padded, %pad, %copy_back = transform.structured.pad %0 pad_to_multiple_of [%c2, 2, 1] {113      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],114      padding_dimensions=[0, 1, 2],115      nofold_flags=[1, 1, 0]116    } : (!transform.any_op, !transform.param<i64>) -> (!transform.any_op, !transform.any_op, !transform.any_op)117    transform.yield118  }119}120 121// -----122 123#map = affine_map<()[s0] -> (-s0 + 12, 7)>124 125// CHECK-LABEL: @static_sizes_output_divisible_on_empty_op126func.func @static_sizes_output_divisible_on_empty_op(%arg0: tensor<24x12xf32>,127    %arg1: tensor<12x25xf32>, %arg2: tensor<24x25xf32>, %iv0: index,128    %iv1: index, %iv2: index) -> tensor<24x25xf32> {129  %0 = affine.min #map()[%iv2]130 131  //      CHECK: %[[T0:.*]] = tensor.empty132  //      CHECK: %[[T1:.*]] = tensor.empty133  //      CHECK: %[[T2:.*]] = tensor.empty134  %1 = tensor.empty(%0) : tensor<4x?xf32>135  %2 = tensor.empty(%0) : tensor<?x5xf32>136  %3 = tensor.empty() : tensor<4x5xf32>137 138  //  CHECK-DAG: %[[CST:.*]] = arith.constant 0.139 140  //      CHECK: %[[T3:.*]] = tensor.pad %[[T0]] nofold141  //      CHECK: tensor.yield %[[CST]]142  //      CHECK: %[[T4:.*]] = tensor.pad %[[T1]] nofold143 144  //      CHECK: %[[T5:.*]] = linalg.matmul145  // CHECK-SAME:              ins(%[[T3]], %[[T4]] : tensor<4x7xf32>, tensor<7x5xf32>)146  // CHECK-SAME:              outs(%[[T2]] : tensor<4x5xf32>)147  %4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>148  %5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>149  func.return %5 : tensor<24x25xf32>150}151 152module attributes {transform.with_named_sequence} {153  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {154    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op155    %padded, %pad, %copy_back = transform.structured.pad %0 {156      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],157      padding_dimensions=[0, 1, 2],158      nofold_flags=[1, 1, 0]159    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)160    transform.yield161  }162}163 164// -----165 166func.func @pad(%arg0: tensor<24x12xf32>,167               %arg1: tensor<12x25xf32>,168               %arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {169  // expected-note @below {{when applied to this op}}170  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>171  func.return %0 : tensor<24x25xf32>172}173 174module attributes {transform.with_named_sequence} {175  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {176    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op177    // expected-error @below {{op expects a padding value of type 'f32', got 0 : i32}}178    %padded, %pad, %copy_back = transform.structured.pad %0 {179      padding_values=[0: i32, 0.0 : f32, 0.0 : f32],180      padding_dimensions=[0, 1, 2],181      nofold_flags=[1, 1, 0]182    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)183    transform.yield184  }185}186 187// -----188 189func.func @pad(%arg0: tensor<24x12xf32>,190               %arg1: tensor<12x25xf32>,191               %arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {192  // expected-note @below {{when applied to this op}}193  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>194  func.return %0 : tensor<24x25xf32>195}196 197module attributes {transform.with_named_sequence} {198  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {199    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op200    // expected-error @below {{expects a padding that parses to 'f32', got "{foo}"}}201    %padded, %pad, %copy_back = transform.structured.pad %0 {202      padding_values=["{foo}", 0.0 : f32, 0.0 : f32],203      padding_dimensions=[0, 1, 2],204      nofold_flags=[1, 1, 0]205    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)206    transform.yield207  }208}209 210// -----211 212// With all padded being static, there's nothing to pad. However, with the213// `nofold` attribute set (see `nofold_flags`), the corresponding pad Ops are214// preserved.215 216// CHECK-LABEL: @zero_pad_static(217func.func @zero_pad_static(%arg0: tensor<24x12xf32>,218                           %arg1: tensor<12x25xf32>,219                           %arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {220 221// CHECK-SAME:      %[[ARG_0:.*]]: tensor<24x12xf32>,222// CHECK-SAME:      %[[ARG_1:.*]]: tensor<12x25xf32>,223// CHECK-SAME:      %[[ARG_2:.*]]: tensor<24x25xf32>) -> tensor<24x25xf32> {224 225// CHECK:           %[[PAD_ARG_0:.*]] = tensor.pad %[[ARG_0]] nofold low[0, 0] high[0, 0]226// CHECK:           %[[PAD_ARG_1:.*]] = tensor.pad %[[ARG_1]] nofold low[0, 0] high[0, 0]227// CHECK-NOT:       tensor.pad228 229// CHECK:           %[[MATMUL:.*]] = linalg.matmul230// CHECK-SAME:      ins(%[[PAD_ARG_0:.*]], %[[PAD_ARG_1:.*]] : tensor<24x12xf32>, tensor<12x25xf32>)231// CHECK-SAME:      outs(%[[ARG_2]]232  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>233  func.return %0 : tensor<24x25xf32>234}235 236module attributes {transform.with_named_sequence} {237  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {238    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op239    %padded, %pad, %copy_back = transform.structured.pad %0 {240      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],241      padding_dimensions=[0, 1, 2],242      nofold_flags=[1, 1, 0]243    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)244    transform.yield245  }246}247 248// -----249 250// With all padded dims being static, there's nothing to pad. However, with the251// `nofold` attribute set (see `nofold_flags`), the corresponding pad Ops are252// preserved. Same as above, but some dims are now dynamic.253 254// CHECK-LABEL: @zero_pad_dynamic(255func.func @zero_pad_dynamic(%arg0: tensor<?x12xf32>,256                            %arg1: tensor<12x?xf32>,257                            %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {258 259// CHECK-SAME:      %[[ARG_0:.*]]: tensor<?x12xf32>,260// CHECK-SAME:      %[[ARG_1:.*]]: tensor<12x?xf32>,261// CHECK-SAME:      %[[ARG_2:.*]]: tensor<?x?xf32>) -> tensor<?x?xf32> {262 263// CHECK:           %[[PAD_ARG_0:.*]] = tensor.pad %[[ARG_0]] nofold low[0, 0] high[0, 0]264// CHECK:           %[[PAD_ARG_1:.*]] = tensor.pad %[[ARG_1]] nofold low[0, 0] high[0, 0]265// CHECK:           %[[PAD_ARG_2:.*]] = tensor.pad %[[ARG_2]] nofold low[0, 0] high[0, 0]266 267// CHECK:           %[[MATMUL:.*]] = linalg.matmul268// CHECK-SAME:      ins(%[[PAD_ARG_0:.*]], %[[PAD_ARG_1:.*]] : tensor<?x12xf32>, tensor<12x?xf32>)269// CHECK-SAME:      outs(%[[PAD_ARG_2]]270  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x12xf32>, tensor<12x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>271  func.return %0 : tensor<?x?xf32>272}273 274module attributes {transform.with_named_sequence} {275  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {276    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op277    %padded, %pad, %copy_back = transform.structured.pad %0 {278      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],279      // Note - only the static dim is padded280      padding_dimensions=[2],281      nofold_flags=[1, 1, 1]282    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)283    transform.yield284  }285}286 287// -----288 289// Impossible to get a bound for padding - fails290 291func.func @negative_no_ub_estimate(%arg0: tensor<?x12xf32>,292                                   %arg1: tensor<12x?xf32>,293                                   %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {294 295  // expected-note @below {{target op}}296  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x12xf32>, tensor<12x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>297  func.return %0 : tensor<?x?xf32>298}299 300module attributes {transform.with_named_sequence} {301  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {302    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op303    // expected-error @below {{failed to pad op}}304    %padded, %pad, %copy_back = transform.structured.pad %0 {305      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],306      // Note - attempting to pad non-static dim307      padding_dimensions=[1],308      nofold_flags=[1, 1, 1]309    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)310    transform.yield311  }312}313 314// -----315 316// Test dynamic padding using `use_prescribed_tensor_shapes`317 318// CHECK: #[[MAP:.*]] = affine_map<()[s0] -> (-s0 + (s0 ceildiv 7) * 7)>319// CHECK: @use_prescribed_tensor_shapes320// CHECK: (%[[ARG0:.*]]: tensor<?x12xf32>, %[[ARG1:.*]]: tensor<12x?xf32>321func.func @use_prescribed_tensor_shapes(%arg0: tensor<?x12xf32>,322                                   %arg1: tensor<12x?xf32>,323                                   %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {324  // CHECK: %[[C1_0:.*]] = arith.constant 1 : index325  // CHECK: %[[DIM_0:.*]] = tensor.dim %[[ARG1]], %[[C1_0]] : tensor<12x?xf32>326  // CHECK: %[[PADDING:.*]] = affine.apply #[[MAP]]()[%[[DIM_0]]]327  // CHECK: %[[PADDED:.*]] = tensor.pad %[[ARG1]] low[0, 0] high[0, %[[PADDING]]] {328  // CHECK: linalg.matmul ins(%[[ARG0]], %[[PADDED]] : tensor<?x12xf32>, tensor<12x?xf32>) 329  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x12xf32>, tensor<12x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>330  func.return %0 : tensor<?x?xf32>331}332 333module attributes {transform.with_named_sequence} {334  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {335    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op336    %padded, %pad, %copy_back = transform.structured.pad %0337    pad_to_multiple_of [7] use_prescribed_tensor_shapes {338      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],339      padding_dimensions=[1]340    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)341    %func = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op342    transform.apply_patterns to %func {343      transform.apply_patterns.canonicalization344    } {apply_cse} : !transform.any_op345    transform.yield346  }347}348 349// -----350 351// Check that the padding can be applied even when the output argument of the352// linalg op is not produced by an empty op or an extract_slice op.353 354// CHECK-DAG: #[[$MAP_MIN:.*]] = affine_map<(d0) -> (-d0 + 2044, 16)>355// CHECK-DAG: #[[$MAP_TO_16:.*]] = affine_map<(d0) -> (-d0 + 16)>356// CHECK-LABEL: @outs_not_produced_by_empty_or_extract_slice(357// CHECK-SAME: %[[A:[^: ]*]]: tensor<128x2044xf32>,358// CHECK-SAME: %[[B:[^: ]*]]: tensor<2044x128xf32>)359func.func @outs_not_produced_by_empty_or_extract_slice(%a : tensor<128x2044xf32>, %b : tensor<2044x128xf32>) -> tensor<128x128xf32> {360  %cst = arith.constant 0.000000e+00 : f32361  %0 = tensor.empty() : tensor<128x128xf32>362  %9 = linalg.fill ins(%cst : f32) outs(%0 : tensor<128x128xf32>) -> tensor<128x128xf32>363 364  %c0 = arith.constant 0 : index365  %c16 = arith.constant 16 : index366  %c2044 = arith.constant 2044 : index367  // CHECK: scf.for %[[ARG3:.*]] = {{.*}} iter_args(%[[ARG4:.*]] = %{{.*}})368  %10 = scf.for %arg3 = %c0 to %c2044 step %c16 iter_args(%arg4 = %9) -> (tensor<128x128xf32>) {369    // CHECK: %[[MIN:.*]] = affine.min #[[$MAP_MIN]](%[[ARG3]])370    %11 = affine.min affine_map<(d0) -> (-d0 + 2044, 16)>(%arg3)371    // CHECK: %[[A_SLICE:.*]] = tensor.extract_slice %[[A]]372    // CHECK: %[[B_SLICE:.*]] = tensor.extract_slice %[[B]]373    %extracted_slice_2 = tensor.extract_slice %a[0, %arg3] [128, %11] [1, 1] : tensor<128x2044xf32> to tensor<128x?xf32>374    %extracted_slice_3 = tensor.extract_slice %b[%arg3, 0] [%11, 128] [1, 1] : tensor<2044x128xf32> to tensor<?x128xf32>375    // CHECK-DAG: %[[CST:.*]] = arith.constant 0.376 377    // CHECK-DAG: %[[TO_16:.*]] = affine.apply #[[$MAP_TO_16]](%[[MIN]])378    // CHECK: %[[PADDED_A_SLICE:.*]] = tensor.pad %[[A_SLICE]] nofold low[0, 0] high[0, %[[TO_16]]]379    // CHECK: tensor.yield %[[CST]]380    // CHECK: %[[PADDED_B_SLICE:.*]] = tensor.pad %[[B_SLICE]] nofold381    // The output shape is already padded, so actually we shouldn't382    // add anything to the upper bound.383    // CHECK: %[[PADDED_ARG4:.*]] = tensor.pad %[[ARG4]] nofold low[{{.*}}] high[0, 0]384 385    //      CHECK: %[[T5:.*]] = linalg.matmul386    // CHECK-SAME:              ins(%[[PADDED_A_SLICE]], %[[PADDED_B_SLICE]] : tensor<128x16xf32>, tensor<16x128xf32>)387    // CHECK-SAME:              outs(%[[PADDED_ARG4]] : tensor<128x128xf32>)388    %res = linalg.matmul ins(%extracted_slice_2, %extracted_slice_3 : tensor<128x?xf32>, tensor<?x128xf32>) outs(%arg4 : tensor<128x128xf32>) -> tensor<128x128xf32>389    scf.yield %res : tensor<128x128xf32>390  }391  return %10 : tensor<128x128xf32>392}393 394module attributes {transform.with_named_sequence} {395  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {396    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op397    %padded, %pad, %copy_back = transform.structured.pad %0 {398      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],399      padding_dimensions=[0, 1, 2],400      nofold_flags=[1, 1, 1]401    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)402    transform.yield403  }404}405 406// -----407 408#map = affine_map<()[s0] -> (-s0 + 12, 7)>409 410// CHECK-LABEL: @pack_everything411func.func @pack_everything(%arg0: tensor<24x12xf32>,412                           %arg1: tensor<12x25xf32>,413                           %arg2: tensor<24x25xf32>,414                           %iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {415  %0 = affine.min #map()[%iv2]416 417  //      CHECK: %[[T0:.*]] = tensor.extract_slice %418  //      CHECK: %[[T1:.*]] = tensor.extract_slice %419  //      CHECK: %[[T2:.*]] = tensor.extract_slice %420  %1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>421  %2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>422  %3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>423 424  //  CHECK-DAG: %[[CST:.*]] = arith.constant 0.425 426  //      CHECK: %[[PAD0:.*]] = tensor.pad %[[T0]] nofold427  //      CHECK: %[[PAD1:.*]] = tensor.pad %[[T1]] nofold428  //      CHECK: %[[PAD2:.*]] = tensor.pad %[[T2]] nofold429 430  //      CHECK: %[[T5:.*]] = linalg.matmul431  // CHECK-SAME:              ins(%[[PAD0]], %[[PAD1]] : tensor<4x7xf32>, tensor<7x5xf32>)432  // CHECK-SAME:              outs(%[[PAD2]] : tensor<4x5xf32>)433 434  // Get unpadded result (no-op in this example).435  //      CHECK: %[[T6:.*]] = tensor.extract_slice %[[T5]]436  // Copy back result to the original buffer, so that the destination of the437  // computation does not change.438  //      CHECK: %[[T7:.*]] = bufferization.materialize_in_destination %[[T6]] in %[[T2]]439  %4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>440 441  //      CHECK: %[[T8:.*]] = tensor.insert_slice %[[T7]] into %{{.*}}442  %5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>443  func.return %5 : tensor<24x25xf32>444}445 446module attributes {transform.with_named_sequence} {447  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {448    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op449    %padded, %pad, %copy_back = transform.structured.pad %0 {450      padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],451      padding_dimensions=[0, 1, 2],452      nofold_flags=[1, 1, 1]453    } : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)454    transform.yield 455  }456}457 458// -----459 460// This test checks that by using `simplify_min_max_affine_ops` after padding461// and tiling, it's possible to recover static tiled slices.462 463// CHECK-LABEL: @dyn_pad_tiling464// CHECK: %[[LHS:.*]] = tensor.pad465// CHECK: %[[RHS:.*]] = tensor.pad466// CHECK: scf.for467// CHECK-DAG: tensor.extract_slice %[[LHS]][0, %{{.*}}] [%{{.*}}, 32]468// CHECK-DAG: tensor.extract_slice %[[RHS]][%{{.*}}, 0] [32, %{{.*}}]469func.func @dyn_pad_tiling(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>, %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {470  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>471  return %0 : tensor<?x?xf32>472}473module attributes {transform.with_named_sequence} {474  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {475    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg0 : (!transform.any_op) -> !transform.any_op476    %padded, %pad, %copy = transform.structured.pad %0 pad_to_multiple_of [32] use_prescribed_tensor_shapes {padding_dimensions = [2], padding_values = [0.000000e+00 : f32, 0.000000e+00 : f32, 0.000000e+00 : f32]} : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)477    %tiled_linalg_op, %loops = transform.structured.tile_using_for %padded tile_sizes [0, 0, 32] : (!transform.any_op) -> (!transform.any_op, !transform.any_op)478    %1 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op479    %2 = transform.apply_registered_pass "resolve-shaped-type-result-dims" to %1 : (!transform.any_op) -> !transform.any_op480    transform.apply_patterns to %2 {481      transform.apply_patterns.canonicalization482    } {apply_cse} : !transform.any_op483    %3 = transform.structured.match ops{["affine.min", "affine.max"]} in %arg0 : (!transform.any_op) -> !transform.any_op484    transform.affine.simplify_min_max_affine_ops %3 : !transform.any_op485    transform.apply_patterns to %2 {486      transform.apply_patterns.canonicalization487    } {apply_cse} : !transform.any_op488    transform.yield 489  }490}491 492